Infrastructure Splits to Serve Humans and Agents Simultaneously

AI assistants answer every question fluently, but human judgment requires knowing when to withhold an answer. Does AI access erode people's willingness to say "I don't know"?

Method: Across five experiments with 3,132 participants, merely having access to AI—even when its advice was engineered to be wrong—nearly eliminated participants' willingness to suspend judgment. They answered more questions but were correct about a third as often as without AI, yet their confidence nearly doubled. Incentivizing accuracy reduced AI reliance and increased suspended judgments, but still far less than when AI was unavailable.

Caveats: Tested on difficult trivia questions. Professional decision contexts with higher stakes may show different patterns.

Reflections: Does the suppression of "I don't know" persist after extended exposure, or do users recalibrate over time? · Can interface design restore suspended judgment without removing AI access entirely? · Do domain experts show the same metacognitive collapse, or does expertise provide resistance?

AI shopping agents must navigate websites designed for humans. Existing SEO and accessibility metrics don't measure whether an agent can reliably extract product details, compare options, or complete purchases.

Method: A controlled experiment with three browser-agent models (GPT-4.1, Gemini-2.5 Flash, Grok-4 Fast) across 300 runs compared identical e-commerce sites—one human-oriented, one agent-ready. The agent-ready version achieved 134 PASS runs versus 74 for the baseline (strict success rates of 89.3% vs. 49.3%), reduced PARTIAL outcomes from 43 to 3, and lowered average step count from 9.31 to 6.49. Gains concentrated in product detail extraction, comparison, and multi-constraint selection.

Caveats: Tested only on e-commerce with controlled catalog and pricing. Other domains and open-ended tasks unverified.

Reflections: Do agent-ready design patterns degrade human usability, or can they coexist? · Which specific structural features drive the largest success rate gains? · How do agent-ready principles transfer to non-transactional domains like content sites or SaaS tools?

AI coding agents read READMEs and install dependencies without verifying package names, sources, or vulnerability status. Setup documentation becomes a vector for supply-chain attacks.

Method: Systematic evaluation across production coding-agent harnesses and frontier models on twelve attack scenarios found agents catch blatant typosquats reliably but miss plausible separator-confusion names (azurecore for azure-core). Source-based attacks like registry redirection are missed almost everywhere—on npm and Cargo, nearly every model installs the untrusted dependency. Security-oriented prompts recover part of the gap only for the dimension they name; the same model catches an attack through one harness and installs it through another.

Caveats: Tested on Python, npm, and Cargo ecosystems. Other package managers and attack vectors not evaluated.

Reflections: Can agents be trained to recognize registry redirection patterns, or is this fundamentally a tooling problem? · Do multi-step verification workflows (e.g., confirm-before-install) reduce attack success without breaking agent autonomy? · How do these vulnerabilities manifest in agents that generate Dockerfiles or CI/CD configs?